Evolved Policy Gradients
OpenAI Blog
Researchers released Evolved Policy Gradients, a method that automatically designs loss functions for training agents rather than using hand-crafted ones. Agents trained with EPG generalized to novel spatial navigation tasks, such as finding objects on opposite sides of rooms from their training experience. This approach enables faster learning on new tasks without requiring manual redesign of training objectives for each new problem.
Why it matters
We’re releasing an experimental metalearning approach called Evolved Policy Gradients, a method that evolves the loss function of learning agents, which can enable fast training on novel tasks. Agents trained with EPG can succeed at basic tasks at test time that were outside their training regime, like learning to navigate to an object on a different side of the room from where it was placed during training.
